Transcription
Chat GPT just released images [music] too, its new image generation model. And today, I'm going to show you what it can actually [music] do. I'll test the best use cases, generate images step-by-step, [music] and show you how I would use it for real creative projects.
Go to the Chat GPT platform and you should have the latest version. As you know, you can click on this plus sign and choose create images, but you can also ask Chat GPT to do that. So, I rarely click on those. What that option has extra is the possibility to choose an art style from a few predefined ones. And for aspect ratio, instead of auto, you can choose a few, but it is not the complete list of ratios there. So, it is still better to mention the ratio you want in the prompt. So, I use Chat GPT normally without that option.
But here is an important part. How good the results are, how well it follows the prompt, and how accurate the image is, it depends on the setting here. If you use instant, it generates faster and does not think as much, but if you use thinking, it will think before generating, so you get better results. You will just consume your limits faster. By the end of the video, I will use only thinking because I have the team plan, which is now called business. But even with that plan, after about 4 hours of generating images, I hit the limit and it puts me on hold for about 10 hours or asks me to get credits if I want to generate more.
I will click on the plus sign. And this time I will add a photo. Since many of you probably want to use your own photo to change something, improve it, or maybe just have fun. So, I have this photo of a woman here that I will use for testing. With that woman attached, I tell Chat GPT that I want the woman to wear an elegant red dress in an exotic location in the same pose, and to keep the face consistent, change the background to fit, and press enter to send it.
Now, because I have thinking on, it will always think about how to do it. And you can even ask it to search online before it generates to find accurate info about a certain subject. Generation can take from half a minute to a few minutes, so it is not as fast as Nano Banana 2. And the result is this one. And I think it keeps the face pretty consistent. So, with the right prompt, you can have good consistency, especially if you keep the pose. It is way better than the previous version of the Chat GPT image generator.
From here, you can download the image. It does not do 2K images like some articles say, maybe internally or with API, but in Chat GPT, it is not 2K. And you can also change the ratio after the image is generated. So, if you got the wrong ratio or want to try a new one, you can do that easily. Let's see how it looks in a 4:3 landscape ratio. Sometimes I like to look at what text it shows there to learn how the AI thinks when it does something. And we got our landscape image that looks pretty good. So, for this use case, it did well, but that does not mean it will work every time. It can still fail sometimes. Sometimes it can get stuck on a subject to keep consistency. So, if it is a new project, always open a new chat, so it starts fresh and does not assume you want something from that chat.
Let's try something else with an illustration this time. So, I want to use this art style to create an illustration for a spaceship. I always mention the ratio in the prompt, so it does not pick a random one, and I can give extra details. So, the original illustration looks like this. It is kind of complex with small lines, but let's see what it can do. So, it captured the style, but the quality of the image is not that good. I noticed that Chat GPT does that randomly for certain art styles, but I am not sure what triggers it. Sometimes with cinematic and fantasy images, it can do that. So, it captured the style, but the quality is bad for some reason. Let me try something else in a new chat.
This time I will add the illustration and the woman, and I want to recreate the portrait of the woman while keeping her identity and pose, but in the style of the mushroom illustration. Make sure it is clear with no artifacts or mistakes. And let's see if we get better results. So, it is thinking again and looking at the images to understand what each one is and how it can integrate them to reach the goal. So, we got our result, and it is better than before. But I think this art style triggered something. Maybe those small triangles in the prompt made the image build from small shapes, and it looks like low quality. Anyway, this rarely happens.
Moving to the next use case. For this one, I want to see if I can copy an art style from Midjourney. If you used it before, they have a styles page, and I can go to top from this month and look around to see if I can find an interesting style that I want to try. So, let's say this style is something I want. I will take a screenshot of all images, so it can see the style in action with different subjects, and paste that screenshot into Chat GPT. Then I ask for a portrait of a bunny character in this art style. The style is quite similar, not identical, but not bad.
Now that it knows the style, let's see if I can add the woman and ask it to adapt this portrait to the same art style while keeping the woman's identity and pose. And we got our result. So, it is the same woman in the same style, but if you are in the same chat, it can take things from that chat. Like I said, it can get stuck on certain images to keep consistency. In this case, it used a similar background to the bunny. If you do not want that to happen, open a new chat and continue there. So, let me repeat the steps quickly. Maybe add some extra words to make sure it changes only the clothes, not the woman. And now we got the results with a random background like I expected in a vintage art style. So, it is worth trying to see what you get.
Let's test the woman again, but this time I want to keep the same pose and identity, and make her look like a businesswoman, and add some makeup. After a minute, we got our woman in a suit with subtle makeup since she is a businesswoman, but it is there around the eyes and lips. So, with a normal photo, you can get a professional photo in a minute. You can rotate it to show different angles, and the result looks similar to the original. It is hard to tell since it is not a real woman, so I do not know how accurate it would be, but it can easily fool me.
Let's see if I can make her older, like an old woman. And it seems it can do it, but we will never know for sure since in 30 to 40 years, we probably will not have Chat GPT to check if it was accurate. But if you look like one of your parents and the result looks like your mother or father, it might be accurate.
Here is another useful case. I took a photo with my phone of a random object I had around, like this gyro ball. Then I asked to replace the background with a solid white background. You can also ask to improve it and color correct it, but I did not want to get that AI look into it. You can easily do this in Photoshop also. AI can just do it as well. So, now it is on a white background, and it created a nice shadow on the ground, and it already looks more professional just by replacing the background. So, you can sell it on online shops or Etsy and so on, and it will attract more buyers.
If you want to go a step further, you can ask it to add some text. I like to keep the same pose and composition since sometimes when you ask for a different angle, it has to invent things and might not be as accurate as the real photo. So, let's say I plan to sell it for $10. By the way, sometimes when it is thinking, you can click on this arrow to see what it does. You can see what it says and what it plans to do. And we got our results, which look pretty good, I would say. It knew what text to put, what color to use, and what the object does. It added some red color to grab attention, and even used a similar font for the title like the one from the actual product. That is pretty this will be very useful for many designers.
I used the pen tool in Illustrator to create the letter A. That took about a minute, and probably AI can do it faster, but I wanted it to be something I made. Now, if I want to create a full font, it would take an hour or more to make all the letters match the same style. Chat GPT is incredibly good at generating the rest of the letters in the same style. So, you can see it started with my letter A and created all the other letters from A to Z without mistakes. Then you can use live trace in Illustrator to get the letters and use an app like Font Shelf to create your own custom font and install it on your computer.
If you do logo design, then you should give Chat GPT a try. I suggest doing a sketch of the logo you want. That way you get a more unique logo design, and it does not have to be perfect. AI can improve it for you. So, what I usually do, I ask for variations of that original sketch to see if I can get something similar that I can use for a real logo. Later, I can make that logo vector and adjust it, or even combine ideas to get a nice logo. On the first try, I got six versions of that logo, and believe me, this saves you a lot of time. I used to spend hours sketching logo ideas. You can also ask it to create another set if you do not like any of those.
So, for the second generation, let's say you pick one that you like, you can push it further and create some mock-up ideas and presentations for it. The presentation is nice, but you can see it used the sketch version instead of the version I asked for. When that happens, it is time for a new chat because it keeps remembering things that it should not, and it does not remember things that it should. I will just take a screenshot of the logo I want and paste it in a new conversation, and then you will get a nice presentation for it, like it was supposed to do in the previous chat, but it acts stupid like it does not understand.
Let's talk a little bit about prompting and how to write clearer prompts so you get better image results. Always start with the goal. Say what the image is for before describing the image itself, because the purpose defines every other choice. Is it a thumbnail, a poster, or a mock-up? Then define the format. Even if it understands the ratio if you put it later. If you put it first, it designs the rest of the prompt according to that ratio, so you get better results. Keep the image focused on one clear purpose, because if you want too many things in one image, it might not work as you expect. Name the main subject clearly and be specific about what element is the hero. Is it someone's face, a burger, or a perfume bottle? Tell the model how you want the scene arranged so you do not get random results. Mention the position, the angle, and so on. Lighting changes the whole image, so use mood words to guide the AI in the right direction. Describe the actual style you want to see. Is it modern, retro, futuristic, or realistic? Now this refers to short text that appears in the image, not the prompt itself. It can handle a brand name or a few words in a poster well, but if you ask for an entire book page, it might make mistakes. So far, it is still the best AI model for text that I tested, but long text can still fail, while short text comes out almost perfect each time. Mention the things you do not want. Add no in front, like no watermark, no duplicate objects, so anything you do not want goes at the end of the prompt.
So the formula for a good prompt is this: purpose, what the image is for, subject, the hero element, style, the design language, composition, the layout and framing, lighting, the atmosphere and mood, text, keep it short, aspect ratio, the size and where you plan to use it, constraints, what to avoid. A strong prompt has a clear purpose, clear structure, and clear limits, and every part of the formula has its own job.
So what is Chat GPT good at? From my tests, it is good at design work and things that contain text, from thumbnails to icons to products and more. You will see more examples in the next minutes. But Chat GPT still has limits like many AI models. Some things are harder to control, like too much text in one image, crowded infographics, or repeating characters exactly. Even if it can keep characters consistent, it is not always perfect. It is also not great at very small details if there are many of them. It can do details well, but I hope in the future they offer real 2K or 4K generation for more detail. So the best prompts are not random words, but clear instructions with a goal, a subject, a look, and limits. Imagine you tell another person how to do something. You would not say, "I want a painting, square, cool, nice." You would give details like size, color, subject, and so on. Better prompts give better results.
I know it is a lot of information, and even if I have used AI for a few years now, I still forget the order sometimes or I am too lazy. That is why I created a custom Chat GPT with that formula, and you can create one, too, but I will share the link in the video description so you can use it for your projects. So in this custom GPT, I usually start with the words "Give me a prompt" so it understands the purpose. Since they added the new model, it sometimes keeps generating images even when you ask only for prompts, as you will see in the next minutes, so I hope they fix that soon. Then I tell it what I want the prompt for. Let's say I want it for a product photo for a brand called Choco, which is a dark 85% chocolate with orange. Maybe I add beautiful product photo, since this is just a prompt for creating a prompt, not an image. The real prompt is generated now for me, a long and detailed prompt using the formula we talked about earlier, so it has clear instructions for everything it needs. Now I can just click on this button to copy. You might ask why do prompts first and then images after. Why not go directly to images? Because images take time, prompts are really fast to generate, and you can see what it plans to do. And second, using a clear formula improves the results of the generated image. So I open a new tab so you can have two windows open, one for chat and one for images, and then I paste the prompt here and start generating.
If you want, you can also generate in the same custom GPT after you get the prompt. I will say generate, and like I told you, even if it is based on a prompt, Chat GPT since the update still generates images for me. So in the second chat, here is the result I got. It designed the chocolate box and the preview for me, so it is not bad. If I check in the custom chat, I got something similar there also. It could be better. Let's do a few more tests. Here I tried another example for a coffee bag for a brand called Toffee, and it looks okay. Now that it created that, I wanted to see if it can think about how the next scene would look with a tribal woman drinking coffee in a way that is related to that product and makes sense. And it added the woman and the coffee, so it might work for advertising different products online.
Now let me try a prompt for a poster. I gave it some details and asked it to invent the rest, and look, it started to generate the image again, even if I asked for a new prompt. Did I make a mistake somewhere or is Chat GPT going crazy? I stopped the generation and say it again that I want a prompt. What the chat? You see, it generates an image again, like it is trying to consume all my limits. Let me try again. Generate a prompt, not an image. And now finally it generates a prompt like it should. So let me try that prompt, and we got a poster. I did not give it much detail, so it is kind of minimalist. Let me ask it to add time, location, and more details so it looks more real, since a real poster usually has more information. So it can do good design when Chat GPT wants, it just has to stress me a bit before it does it. By the way, for a custom GPT, you create a new chat from here using this button, because if you use the other button, it creates a new chat, but not with this custom GPT. Just something to keep in mind. So I start a new chat for each new project to avoid confusion. I mean AI confusion.
I want to create now a set of six stickers, something with a cute panda on a solid white background. Let me copy and generate that. The result is quite good. I like it. Let's say I like the first one and I want only that one. I can take a screenshot of it, then generate a square ratio image of that panda. I will mention high quality because it might improve the quality a bit. I will ask to remove the white contour, since I do not need it in this case, and to give me only the panda on a transparent background. Before, transparent background did not always work correctly or it tried to use tools to remove it, but now I actually got a panda on a transparent background. You can see when I hover over it, I get that transparency grid. So now I can save it, and if I open it on my computer on a different background, you can see it is indeed transparent. So that is a good thing to have.
Let's move to the next use case. This time I want to create a menu, let's say for a fast food restaurant that has only three products, and maybe I add some prices there, American style. I got something, but maybe I adjust the prices to end in 99. And again, it tries to generate images. It is so annoying when I am in a hurry and it starts generating images when it should create prompts. It can be very smart on some things and still mess this up. So I finally got a prompt, and let's test that. We got our menu, it looks more like a flyer, but since it only had three products, that is what it could do. But I am used to bigger menus, so let's actually create a menu with more products for a pizzeria. That should work better with multiple pizzas to look like a proper restaurant menu, and we got a nice prompt. Let's test that. And look at this menu design, how good it looks. So you can see how your prompt and all that information helped for better results. It will not always work, but it is capable of great things when it wants. Now I need to order a pizza.
If you are curious what this custom GPT has in the instructions, I can show you using this edit button. You do not have it, since only the person who created the GPT can edit it. So basically, it has a long list of instructions on how to prompt and in what order, like we talked before with that formula. Then I added more instructions in a text file and put it in knowledge, since the instructions have a limit on how long they can be. For capability, it has a few enabled. I will disable image generation, maybe that way it will generate only prompts, but I doubt it. Let me update it, so now everyone who accesses the link should have this version. Let's test it to see how it goes.
Let me try a fashion magazine layout, something professional. And you can see, even if I turned off image generation, it still generated images instead of a prompt. So, I think it has some bugs since the last update. Let me edit the prompt and ask again for only the prompt. Now, I got the prompt like it should. So annoying. After we test it, we got this magazine cover. I have seen better, but you can try different prompts.
Now, let's say I want the interior of the magazine, like the first two pages design. I am not sure why it used the exact same pose and the same woman on the first page, but I can change that later. I am more interested in the layout and how it arranges things to look good. Then we can continue. Maybe I do pages three and four now and mention that I want jewelry on that page. Let's see what we get. It is still the same woman, but at least in a different pose. And the page on the right looks quite nice. So, you can get some inspiration if you design magazines.
You can try to create all kinds of interfaces, from operating systems to apps, web design pages, or mobile apps. In a few tries, you can get some really nice results and inspiration that you can later code with chat GPT or Claude by prompting. These days, vibe coding has become much easier. And of course, you can create all kinds of product photos and food photography from text to use for magazines, blogs, articles, and anything else you might need. I see some people even using AI images in shops, but I do not recommend that since clients should see how the actual product looks. It is one thing when you edit colors and change the background, and another when it is a completely generated image.
It is quite good at vector and modern illustrations, like this avatar I did for the custom GPT. You can also create flat lay images with different objects, maybe for presentations or if you need to create mock-ups. You can even create photos of random people if you need placeholders for websites or anything to show your client instead of using a basic placeholder. And if you are a YouTuber like me, you can have a lot of uses, from headers to avatars to all kinds of banners and YouTube thumbnails. It is good for cartoon designs also, like here where I wanted that bunny with a melting ice cream, it looks pretty cute. You can try cover art for podcasts or music albums. It is quite good at that. Portfolio images are another use. You can add your own images and ask it to arrange them nicely, since many people have great work but struggle to present it well. Here is an example of a website hero page that looks quite nice. If you are a graphic designer or web designer, you know how much time it takes to create these things in Photoshop. Now, you can get something similar in minutes. It is also good at botanical illustrations. So, if you like that style or you need illustrations for Etsy or products, it is worth trying. Concept art showcase is another good use, especially for concept artists who want inspiration for portfolios and ideas. Children's book illustrations are also worth exploring. You can create full books or coloring books. I have seen many online, and now you can do it better without the common mistakes. Comic book pages or anime pages are something you might want to try. I do not have much experience with those, but it should be able to handle different styles and even full scenes that make sense. Watercolor, pastel, or other painting styles also work well. It knows a lot of art styles, and if you are not sure how to describe one, you can give it a few images in that style and ask it to follow them. For those working with luxury and beauty products, chat GPT is very good. It can create really nice visuals for that. It is also quite good at logo presentations, like here where it created the logo and the mock-up, and it looks quite professional.
After I finish the video, I will create a YouTube thumbnail for my video. I gave it some screenshots of my YouTube page so it can match the style, and then I will adjust it to get a nice result. Maybe I will use something similar to this one. I will see.
That is all for today. I want to thank AI Titans for the support and everyone who subscribed to the membership. If you found something useful, please leave a like and a comment to help the channel. Have a great day.